Bayesian Software Reliability Models Based on Martingale Processes
Sanjib Basu, Nader Baradaran Ebrahimi · Technometrics · 2003
This article proposes new Bayesian models for software reliability based on a piecewise constant failure rate. A martingale process prior is assumed on the failure rate. Three different hyperprior models on the martingale process are considered. The first two models assume that the conditional variance is dependent on the mean, whereas the third model assumes constant conditional variance. Markov chain sampling–based posterior analysis and prequential predictions are developed for these models and are illustrated in a software failure dataset. In addition, model comparison is studied via the Bayes factor criterion. General techniques are described for estimating the marginal likelihood of the proposed models as well as of many existing software reliability models, and these are illustrated in two datasets.